Cross‐Exposure Transferability and Failure Boundaries of Self‐Supervised DNN Denoising for AR‐HAXPES Depth Profiling

ABSTRACT Laboratory‐source angle‐resolved hard X‐ray photoelectron spectroscopy (AR‐HAXPES) is photon‐starved at the spatial resolution required for mapping, where the frame averaging needed to recover a clean depth profile at every position drives total acquisition time into a prohibitive regime. Self‐supervised deep‐neural‐network (DNN) denoising offers an alternative to acquisition‐time scaling, and a recent companion study has systematically benchmarked its performance against alternative strategies; however, its operational envelope across exposures, its cross‐exposure transferability, and its failure modes in the photon‐starved regime remain uncharacterized. Here, we evaluate self‐supervised denoising at the smallest possible target window across three exposures spanning a 25‐fold range of per‐frame photon flux on a model C/Al 2 O 3 /TiO 2 /Si multilayer. We find cross‐exposure denoiser deployment to agree with an empirical directional rule—train signal‐to‐noise ratio ( S / N ) ≥ inference S / N —characterize the failure boundary in the photon‐starved regime where upstream denoiser collapse propagates into constraint‐saturated inversion solutions, and locate the asymptotic reach of all tested denoising strategies (frame averaging, Self‐DNN, Cross‐DNN) relative to a methodological floor set by the reproducibility of the reference. Within the success region of this rule, cross‐exposure transfer enables up to a 48‐fold equivalent acquisition‐time reduction on the depth root‐mean‐square error (RMSE) axis (14‐fold on the Wasserstein‐1 interface‐localization axis) at the most photon‐starved 4.8 s/frame condition. These results frame self‐supervised AR‐HAXPES denoising as a characterizable engineering operation with quantifiable failure boundaries and recovery routes, rather than as an empirical post‐processing recipe.